Skip to content

Data Handling

All data processing happens inside the consumer’s environment. On Snowflake, this means the consumer’s Snowflake account. On Databricks, the consumer’s workspace. No data is sent to Integrated Quantum Technologies or any external service.

DataAccessPurpose
Feature columns (training)SELECT via warehouseRead consumer-specified columns to train encoder
Feature columns (inference)Masking policy interceptEncode values at query time
Model artifactsInternal stage read/writeStore and load ONNX encoder files
Encoder registryApp-owned tablesTrack encoder metadata, training jobs, configuration
  • No consumer credentials are requested or accepted
  • No data is sent to the provider account
  • No data is sent to external APIs or services
  • No telemetry, analytics, or usage data leaves the consumer environment (billing events are emitted to Snowflake’s billing system, not to the provider)

Non-invertible. The encoder-only half of the autoencoder is exported. The decoder is discarded after training. Exact reconstruction of raw input from the latent vector is not possible.

Task-relevant information. If the latent vector is useful for ML, it contains information correlated with the task. When sensitive attributes correlate with the task, partial information about those attributes is present in the encoding. Non-invertible does not mean non-informative.

Deterministic. The same input always produces the same output. This preserves joins, aggregations, and downstream pipeline reproducibility.

EventWhat happens
TrainingRaw data is read via SELECT, processed in-memory, never persisted outside the ONNX model
InferenceRaw values enter the encoding function, latent vectors are returned. Raw values are not logged or stored.
Model deletionActive model file, all archived versions, feature view, feature function, and registry entry are removed
App uninstallAll app-owned objects (schemas, stages, compute pools) are dropped by Snowflake

AIQu VEIL uses a deterministic autoencoder architecture (PyTorch training, ONNX Runtime inference). The model is trained exclusively on consumer data. No pre-trained models from external sources are used. No foundation models or generative AI components are involved.

The full ML/AI disclosure is available at VEIL ML/AI Disclosure.